Estimation of individual tree metrics using structure-from-motion photogrammetry.

(1)

Estimation of Individual Tree Metrics using

Structure-from-Motion Photogrammetry

A thesis submitted in partial fulfilment of the requirements

for the Degree of Master of Science in Geography

Jordan M. Miller

Department of Geography

University of Canterbury

(2)

I

Abstract

The deficiencies of traditional dendrometry mean improvements in methods of tree mensuration are necessary in order to obtain accurate tree metrics for applications such as resource appraisal, and biophysical and ecological modelling. This thesis tests the potential of SfM-MVS (Structure-from-Motion with Multi-View Stereo-photogrammetry) using the software package PhotoScan Professional, for accurately determining linear (2D) and volumetric (3D) tree metrics. SfM is a remote sensing technique, in which the 3D position of objects is calculated from a series of photographs, resulting in a 3D point cloud model. Unlike other photogrammetric techniques, SfM requires no control points or camera calibration. The MVS component of model reconstruction generates a mesh surface based on the structure of the SfM point cloud.

The study was divided into two research components, for which two different groups of study trees were used: 1) 30 small, potted ‘nursery’ trees (mean height 2.98 m), for which exact measurements could be made and field settings could be modified, and; 2) 35 mature ‘landscape’ trees (mean height 8.6 m) located in parks and reserves in urban areas around the South Island, New Zealand, for which field settings could not be modified.

The first component of research tested the ability of SfM-MVS to reconstruct spatially-accurate 3D models from which 2D (height, crown spread, crown depth, stem diameter) and 3D (volume) tree metrics could be estimated. Each of the 30 nursery trees was photographed and measured with traditional dendrometry to obtain ground truth values with which to evaluate against SfM-MVS estimates. The trees were destructively sampled by way of xylometry, in order to obtain true volume values. The RMSE for SfM-MVS estimates of linear tree metrics ranged between 2.6% and 20.7%, and between 12.3% and 47.5% for volumetric tree metrics. Tree stems were reconstructed very well though slender stems and branches were reconstructed poorly.

The second component of research tested the ability of SfM-MVS to reconstruct spatially-accurate 3D models from which height and DBH could be estimated. Each of the 35 landscape trees, which varied in height and species, were photographed, and ground truth values were obtained to evaluate against SfM-MVS estimates. As well as this, each photoset was thinned to find the minimum number of images required to achieve total image alignment in PhotoScan and produce an SfM point cloud (minimum photoset), from which 2D metrics could be estimated. The height and DBH were estimated by SfM-MVS from the complete photosets with RMSE of 6.2% and 5.6% respectively. The height and DBH were estimated from the minimum photosets with RMSE of 9.3% and 7.4% respectively. The minimum number of images required to achieve total alignment was between 20 and 50. There does not appear to be a correlation between the minimum number of images required for alignment and the error in the estimates of height or DBH (R2=0.001 and 0.09 respectively). Tree height does not appear to affect the minimum number of images required for image alignment (R2=0.08).

(3)

II External parameters, which include things such as the position of the tree relative to its surroundings, the background scene and the ambient lighting, appear to be the primary determinants of model success.

The results show that SfM-MVS is capable of producing estimates of 3D and 2D metrics with accuracy at least equal to that of laser scanning and potentially much more accurate than allometric models and traditional dendrometry techniques. SfM-MVS provides a low-cost alternative to remote sensing technologies currently used such as terrestrial laser scanning, and as no specialised equipment is required it is able to be used by people with little expertise or training. Future research is required for exploring the suitability of SfM-MVS for specific applications requiring accurate dendrometry.

(4)

III

Table of Figures

Figure 3.1: Examples of some of the trees used in the nursery tree study ... 22

Figure 3.2: The red tape used to mark the measurement locations on the study trees ... 24

Figure 3.3: The four stages of model generation in PhotoScan. ... 26

Figure 3.4: The blue squares in PhotoScan indicating locations from which photos were taken ... 27

Figure 3.5: Examples of scale bars. ... 29

Figure 3.6: Measuring nursery tree volume. ... 30

Figure 3.7: Mesh surface model. ... 31

Figure 3.8: Regression analysis between SfM-MVS volume estimates and ground truth data. ... 35

Figure 3.9: Regression analysis of SfM-MVS height and crown depth estimates and ground truth data ... 36

Figure 3.10: Regression analysis of SfM-MVS crown spread estimates and ground truth data ... 37

Figure 3.11: Regression analysis of SfM-MVS DBH and combined stem diameters estimates and ground truth data ... 38

Figure 3.12: Relative and absolute esidual plots for SfM-MVS combined stem diameter estimates and ground truth data ... 39

Figure 3.12: High quality SfM-MVS reconstruction ... 46

Figure 4.1: Examples of some of the trees used in the landscape tree study. ... 51

Figure 4.2: Examples of the landscape trees dense point cloud models.. ... 54

Figure 4.3: Examples of the landscape trees dense point cloud models. ... 55

Figure 4.4: The surveying poles that were used as scaling devices ... 56

Figure 4.5: Regression analysis between the size of the minimum photoset and the mean height error and DBH error in SfM-MVS estimates. ... 59

Figure 4.6: Regression analysis between the size of the minimum photoset required for image alignment and the mean tree height for each respective minimum photoset size ... 60

Figure 4.7: Regression analysis between ground truth height values and SfM-MVS height estimates for the complete and minimum photosets. ... 61

Figure 4.8: Regression between ground truth DBH values and SfM-MVS DBH estimates for the complete and minimum photosets. ... 62

Figure 4.9: High quality tree stem reconstruction ... 67

Figure 5.1: The effect of bright sunlight on model reconstruction... 73

Figure 5.2: Poorly reconstructed slender branches and the value of the red tape. ... 75

(8)

VII

Figure 5.4: Poorly reconstructed slender branches . ... 78

Figure 5.5: The close holes tool in PhotoScan ... 79

Figure 5.6: Two examples of trees with excurrent branch growth. ... 83

Figure 5.7: Two examples of trees with decurrent branch growth ... 84

(9)

VIII

Table of Tables

Table 2.1: Accuracies of height estimates achieved in previous TLS studies. ... 10 Table 2.2: Accuracies of DBH estimates achieved in previous TLS studies ... 10 Table 2.3: Accuracies of volume estimates achieved in previous TLS studies. ... 11 Table 2.4: Accuracies of estimates of tree metrics achieved in previous photogrammetric studies. .. 13 Table 2.5: Accuracies of height and biomass estimates achieved in previous ASfM studies ... 17 Table 2.6: Accuracies of height and stem diameter estimates achieved in previous TSfM studies ... 18 Table 3.1 Descriptive statistics for tree volume determined via SfM-MVS and ground truth techniques ... 33 Table 3.2 RMSE and bias statistics for SfM-MVS volume estimates. ... 33 Table 3.3 Descriptive statistics for tree height and crown spread determined via SfM-MVS and ground truth techniques. ... 36 Table 3.4 RMSE and bias statistics for SfM-MVS height and crown estimates. ... 36 Table 3.5 Descriptive statistics for stem diameters determined via SfM-MVS and ground truth techniques. ... 38 Table 3.6 RMSE and bias statistics for SfM-MVS stem diameter estimates. ... 38 Table 4.1: Descriptive statistics for height and DBH from the complete photoset and minimum photosets determined via SfM-MVS and ground truth techniques. ... 57 Table 4.2: The number of images required for image alignment and the % error associated with the height estimates of each ... 58 Table 4.3: The number of images required for image alignment and the % error associated with the DBH estimates of each ... 58 Table 4.4 The number of images required for image alignment and the mean tree height associated with each ... 59 Table 4.4: RMSE and bias statistics for SfM-MVS height and DBH estimates for the complete and minimum photosets. ... 61 Table 4.6: The approximate times for model reconstruction using different size photosets ... 63 Table 7.1: Geographic co-ordinates for each of the landscape study trees………63

(10)

IX

Acknowledgements

I would like to sincerely thank Dr. Justin Morgenroth and Dr. Chris Gomez for introducing me to this area of research, and for their guidance, direction and advice through every stage of this thesis. I am grateful for funding provided by the TREE Fund’s John Z. Duling research grant. Thank you to Lachlan Kirk for the help with field work and preparation for tree volume measurement, and thank you to Mike Smith and Joe Cartman at Christchurch City Council’s Harewood Nursery for their help setting up the nursery trees for use. Finally, thank you to my family and friends for their great help with field work and editing, and their support and encouragement throughout the completion of this thesis

(11)

1

Chapter One

Introduction

1.1 Background

The ability to obtain accurate tree metrics is important for a variety of applications within the fields of commercial forestry (Næsset, 1997; Clark et al., 2000b; Henning & Radtke, 2006), urban forestry (Moskal & Zheng, 2012; Tanhuanpää et al., 2013; Saarinen et al., 2014), ecology (Dandois & Ellis, 2010; Liang et al., 2014b) and horticulture (Rosell et al., 2009). Dendrometry, which is the measure of tree structure and dimension (Morgenroth & Gomez, 2014) is fundamental for resource appraisal (e.g. wood volume, biomass and tree growth) and analysing forest structure (Næsset, 2002, Skowronski et al., 2014) particularly for commercial forest management. Biophysical and ecological information to be derived from tree metrics, include carbon stocks (Houghton, 2005; Strahler et al., 2008), biofuel potential (Dassot et al., 2011; Kankare et al., 2013a) and fire spread risk (Dandois & Ellis, 2010), habitat size and quality (Moskal & Zheng, 2012), storm water attenuation (Xiao & McPherson, 2011) and contribution to reduction of urban air pollution (Nowak et al., 2010). Traditional field methods of dendrometry are prone to error, so they do not always adequately measure tree size and architecture. As a consequence, error is inherently introduced into commercial, biophysical and ecological characteristics which rely on accurate mensuration. Structure-from Motion with Multi-View Stereo-photogrammetry (SfM-MVS) is a relatively new photogrammetric approach allowing automated reconstruction of 3D models using sets of overlapping 2D digital images. SfM-MVS has the potential to provide accurate estimations of tree size and architecture, whilst significantly reducing the difficulties, cost and error associated with traditional dendrometry. Though the method has previously been shown to accurately measure the height and diameter of one individual tree accurately (Morgenroth & Gomez, 2014), it is untested for 3D metrics (e.g. volume) and methodological aspects of its use remain in their infancy with no standardised methods developed yet.

1.2 Current Methods for Traditional Dendrometry

Traditional dendrometry involves terrestrial field mensuration of tree metrics and has been widely adopted due to its simplicity and standardisation, though it is partly due to this simplicity that it is inherently flawed (Clark et al., 2000b; Hopkinson et al., 2004; Morgenroth & Gomez, 2014). Traditional dendrometry is error-prone, laborious, time consuming, expensive (Hopkinson et al., 2004; Melkas et al., 2008; Watt & Donoghue, 2005) and sometimes accessibility to conduct them can be limited by dense vegetation, rough terrain or remoteness (Næsset, 1997; Clark et al., 2000b; Li et al., 2012). Tree height and stem diameter (diameter at breast height, DBH) are among the most important metrics in commercial forest management (Elzinga et al., 2005; Kitihara et al., 2010; Vastaranta et al., 2013) as they provide the means to allometrically estimate the volume of wood, which is the principle commercial product of forestry (West, 2009) and the world’s largest source of renewable construction material.

(12)

2 Field mensuration of tree height is a challenging task, especially for inexperienced observers (Kitahara et al., 2010). At present hypsometers are commonly used (Kitahara et al., 2010; Liang et al., 2014b). These combine clinometers and laser rangefinders and use geometric and trigonometric functions to estimate tree height (West, 2009). Hypsometers are subject to error as they assume that all angles and distances have been measured correctly and that the observer has correctly identified the tallest part of the tree (Morgenroth & Gomez, 2014). Errors can also occur when observers fail to consider ground-slope, tree lean or crown shape (Bragg, 2008). Height is difficult to measure in closed-canopy forests (which the majority of commercial forest plantations are), where the tops of trees are often not visible (Chave et al., 2005). Small trees can be measured with height poles, but the utility of this tool is limited with tall trees. Error in tree height and DBH measurements can be reduced with training but it cannot be totally eliminated, even in experienced observers. DBH is typically measured with callipers or a diameter tape, depending on the thickness of the stem. While the name suggests a somewhat arbitrary height, DBH is most commonly measured at 1.3 m above the ground in literature (Brokaw & Thompson, 2000), and 1.4 m above the ground in New Zealand. DBH provides a repeatable measure that can be applied to almost any tree, though error in its measurement can arise from a number of sources; placement of the tape above or below the recorded height, failure to place the tape perpendicular to the tree axis, or through misreading the tape (Elzinga et al., 2005). Procedural decisions made in forest management, such as thinning and felling are often made directly from DBH measurements; however, the stem form and taper is the key characteristic that determines the yield and quality of the merchantable wood in the stem (Kankare et al., 2013a).

The taper of the stem is fundamental as it depicts the stem diameter as a function of height (Yang et al., 2009) and ultimately estimate stem volume. Stem form cannot be measured in a cost-efficient manner with traditional dendrometry so allometric models are mainly relied upon to describe stem taper; each of these incorporate the height (or length) of the stem, and the diameter at various points along it. Much research has gone into developing these models to describe the stem taper of different tree species in different growing regions, which allow the volume to be derived from DBH and height alone (West, 2009). However, studies have shown that stem tapers vary greatly globally due to different growing conditions and therefore their applicability is limited due to their inaccuracy (Liang et al., 2014a). Stem taper models were developed for commercial forestry to calculate the volume of merchantable timber in tree stems and although they are sufficiently accurate for this purpose, they have little value for tree mensuration in non-commercial forest settings. Perhaps the most obvious limitation of field mensuration is that it can only be conducted at a local scale. Commercial forest plantations often cover thousands of hectares, and the time and labour required dictates that only a small plot sample in a fixed area can be measured (Breidenbach

et al., 2014), typically in the shape of circular plot or transects, as this is more efficient than measuring a random sample of trees. The key assumption in plot sampling is that the trees selected are representative of the whole plantation.

There are modern tools and methods capable of calculating stem volume to a more precise degree (e.g. xylometry and remote sensing) though none of them are always practical nor appropriate for field mensuration of standing trees, and traditional dendrometry is still relied upon for ground-truth validation (West, 2009; Berger et al., 2014). Stem volumes of standing trees derived from allometric equations contain errors due to their often-unrealistic assumptions of natural tree form (Dassot et

(13)

3

al., 2011). As trees in nature never have perfect geometric form (e.g. perfectly cylindrical or conical with no irregularities and an even taper) error is inevitable in stem volume calculations (West, 2009; Morgenroth & Gomez, 2014).

Felling trees offers the most accurate method of volume measurement, as in most cases diameter measurements from the middle and top of the stem are not possible to obtain from standing trees without dangerous climbing techniques, optical dendrometers (Henning & Radtke, 2006; West, 2009) or logging machines (Liang et al., 2014a). However, felling the tree often defeats the purpose of taking the measurement, as the method should in no way cause disruption to the tree’s growth (Clark et al., 2000b). Felling a tree means no future growth can be monitored, not to mention the fact that felling is usually unacceptable in non-commercial forestry settings, such as urban forests or conservations areas (Parker & Matney, 1998; Liang et al., 2014b). For non-destructive field measurements DBH is normally used as the sole diameter for the stem (Melkas et al., 2008; Vastaranta, et al., 2009) and has become the standard height for diameter measurement.

Merchantable stem volume is not the only important commercial product of forestry. There is increasing interest in mensuration of biomass (Repola, 2009; Lin et al., 2010; Skowronski et al., 2014) which is sought for a variety of applications, including carbon accounting and quantification of the amount of organic fuel available for energy production (biofuel). The biomass of a tree comprises of its entire mass, including the stem, branches, bark, foliage and roots. The density of the wood is also considered, as biomass is measured by weight rather than by volume. Biomass is often measured as ‘above-ground’ biomass (AGB) because the roots are difficult to measure. Field estimation of biomass largely involves the same aforementioned methods used for measuring stem volume, and therefore error is inevitable, especially because the crown branches and foliage are much more difficult to measure and large errors can be made (Dassot et al., 2012).

Until recently, remote sensing forest biomass estimates have been carried out with medium resolution satellite imagery and radar, which while efficient in terms of the area they cover, have only had limited success (Houghton, 2005; Kankare et al., 2013a). Currently laser scanning provides the most accurate and efficient method of measurement due to its ability to measure canopy attributes as well as forest structure beneath (Lin et al., 2010) and because biomass is linearly related to volume, estimations can be made from commercial forestry inventories if the data is available (Houghton, 2005; Boudewyn et al., 2007). Data for the latter is primarily based on the volume of merchantable wood (i.e. the emphasis is on the biomass of the stem) (Fraver et al., 2007; Dassot et al., 2012) and while they provide reasonable estimates there is much room for improvement when it comes to monitoring high resolution changes, i.e. at an individual tree scale (Lin et al., 2010; Kankare et al., 2013b). Laser scanning also has limitations that mean it is not always the most appropriate method to use. Biomass models must produce reliable estimates and be applicable at a national scale (Repola, 2009). National and regional biomass estimates would ideally be derived from a probability design sample of ground plots representing a variety of typical forested and vegetated cover types, species composition, age and vertical structure (Boudewyn et al., 2007), but this is not a feasible option considering the high costs that would be involved.

The only way to get a true measure of total biomass is by felling the tree, digging up the roots, and weighing individual components of the tree successively (Lin et al., 2010) or by using xylometry (and converting the volume to mass). This is clearly not feasible in practice (Repola, 2009; Kankare et al.,

(14)

4 2013a), and therefore like stem volume, the development of models to calculate biomass from standard tree metrics has been the subject of several studies over the last few decades (e.g. Marklund, 1988; Repola, 2008; 2009). Regression models for different tree species have been developed from extensive field measurements of standard tree metrics (Houghton, 2005; Zianis et al., 2005; Repola, 2009). Despite this, no species-specific model can be applicable to any tree of a certain species without at least some error. Like stem taper, biomass is specific to particular species and growing regions. For a given species, the biomass of individual trees in a particular plot is more closely correlated to other trees within that plot as opposed to trees of the same species in other plots, showing that local conditions play an important role in tree growth (Repola, 2008).

1.3 The Need for Accurate Tree Mensuration

Commercial forestry inventory is dependent on height and DBH, as well as stem taper and volume. Commercial inventories require frequent gathering of accurate metrics to determine the potential merchantable timber yield, as well as for future modelling and planning of the forest resource. Stem-only harvesting in particular is concerned with the volume of the tree stem, as this contains the largest proportion of a tree’s mass (75-85%, Kankare et al., 2013b) and is the largest continuous section of the tree from which lengths of timber can be cut; therefore it holds the most value in harvesting for various forms of timber products. The roots, small branches and foliage of trees are seldom used as their few potential uses are of little commercial value.

Knowledge of forest biomass is essential for carbon accounting, which has become an integral component of forest mensuration due to rising awareness of the effects of greenhouse gases on the world’s climate (West, 2009). Carbon dioxide (CO2) is the dominant greenhouse gas in the atmosphere and its increase is attributed to combustion of fossil fuels (80%) and global deforestation (Nowak & Crane, 2002). Trees act as a sink for atmospheric CO2 by using it as part of the photosynthesis process and storing it in their tissue. And consequently, a large proportion of the world’s carbon is locked within tree biomass (Nowak & Crane, 2002; Hyyppä et al., 2012; Kankare et al., 2013b) and therein lays the need for biomass quantification. Measuring the carbon sequestered in biomass requires laboratory equipment and cannot be done in the field, though past research has established the typical carbon content for various species of tree, meaning carbon content can be calculated from biomass alone. The carbon content is usually close to 50% of a tree’s biomass and does not vary much between species or different parts of the tree (West, 2009).

There remains uncertainty around the magnitude of carbon uptake by global forests and the distribution of carbon stocks in forest ecosystems (Houghton, 2005; Pan et al., 2011). When calculating carbon flux on a global scale, an accumulation of inaccuracies in biomass calculations will result in significant overall error. The vast boreal forests of Canada, Russia and Scandinavia are estimated to store 30% of global carbon, though research of urban forests has been of particular interest as the latter are subject to the highest concentrations of air pollution due to their proximity to industrial areas, as well as being home to the high densities of people. In 2010 Nowak et al. estimated that in Chicago trees sequester a gross amount of 25,200 tons of carbon annually, with an associated value of $521,000. These same Chicago trees are estimated to hold 716,000 tons in long term storage with a value of $14.8 million. Atlanta and New York City potentially hold even larger stores, with an estimated 1.34 million and 1.35 million tons respectively. Zhao et al. (2010)

(15)

5 estimated that urban forests in the province of Hangzhou, China sequestered 1.33 million tons of carbon annually, accounting for an annual offset of 18.6%. Russo et al. (2014) estimated that urban trees in Bolzano, Italy sequester 12-17 kg of carbon per year

A number of international initiatives have been implemented to encourage countries to decrease rates of carbon emission and offset what they do emit by planting new forests that will act as a sinks. These include the United Nations Collaborative Programme on Reducing Emissions from Deforestation and Forest Degradation in Developing Countries, United National Framework Convention on Climate Change andLand Use, Land Use Change and Forestry section in the Kyoto Protocol (Skowronski et al., 2014). In accordance with the Intergovernmental Panel on Climate Change (2006), countries that plan to mitigate climate change through forestry must provide verified reports (including the uncertainty) on national forest carbon stocks and subsequent contribution to the global carbon cycle and effects on climate change (Næsset & Gobakken, 2008; Berger et al., 2014). The accuracy of biomass mensuration has major implications on the potential for financial compensation which can be achieved through increased offset, as well as introducing bias for future projections (Breidenbach et al., 2014). It is therefore in each country’s best interest to ensure that biomass is calculated to the most accurate degree possible. Biomass at regional through to global scale is changing constantly, particularly because of rapid deforestation that is occurring in tropical regions (Houghton, 2005; Pan et al., 2011). Natural events like cyclones, forest fires and insect swarms can also cause major changes in short duration. It is therefore necessary to conduct continuous or at least very frequent monitoring of biomass in order to know the absolute spatial distribution. This presents a huge challenge considering the necessity for high accuracy (Kankare et al., 2013a).

While the commercial value of wood products has been the major driver behind advances and refinement of dendrometry methods, accurate dendrometry is also important for other applications that do not have as much of a commercial bent. Urban forestry is a growing research discipline regarding the processes and interactions between trees and their surrounding urban environments. Trees and forests in urban landscapes are environmental assets and contribute a range of positive environmental and social effects to the surrounding urban area (Nowak et al., 2010). It is difficult to put a monetary value on these benefits as there is often there is no real way to measure the benefit they have, though some recent efforts have been made (e.g. Nowak et al., 2002; 2010), the more recent of whom developed a valuation system based on environmental, social and economic criteria. Urban forestry includes the management and research of trees in parks, public land and street plantings. The need for information on urban trees is increasing (Holopainen et al., 2013; Saarinen et al., 2014) and inventories are being developed in many cities to manage urban tree populations (Tanhuanpää et al., 2013) which can number in the millions and cover a large proportion of a city’s area. For example, Chicago is estimated to have 3.5 million trees which cover 17.2% of the urban area (Nowak et al., 2010).

Typical tree registries contain location, height and DBH, as well as information on the vitality of trees (Holopainen et al., 2013). As urban trees are not considered for their timber volume, accurate mensuration of some standard tree metrics (e.g. DBH) is not as important as it is in commercial forestry. However, the spatial dimensions of trees may be required for a variety of other urban planning purposes, as well as assessment of ecosystem services, such as storm water attenuation, effects of wind flow, effects on land stability, shading or neighbouring property, air pollution

(16)

6 reduction, fire spread risk and biodiversity assessment. Qualitative data may also be sought to provide information on tree health for maintenance purposes, such as recognizing trees that require watering or pruning and identifying dying trees or trees that have been damaged by strong winds or heavy snowfall may be hazardous to citizens (Tanhuanpää et al., 2013). Interpretation of aerial photography and field mensuration are the conventional methods used to update urban tree registries. Updating tree registries with field mensuration is inefficient, particularly when tens of thousands of trees need to be monitored (Holopainen et al., 2013). In the last decade laser scanning has been commonly used and in many cases datasets for these has already been collected for other urban planning purposes, as well as generating virtual city layouts (Saarinen et al., 2014).

The role of urban trees in mitigating urban pollution has been the subject of much research. Air pollution, which includes gases such as ozone, nitrogen dioxide, carbon monoxide and particulate matter <10 microns in size, is harmful to human health, can cause damage to buildings and ecosystem processes and reduces visibility (Nowak et al., 2010). Trees help to improve urban air quality in a number of ways, including direct removal harmful gases and pollutants through uptake into their leaves, reducing air temperature and lowering the energy consumption of buildings (as well as the subsequent emissions from power stations). They also help by passively capturing airborne particles on their surfaces. Some particles can be absorbed into the tree, though most particles remain on the surface and are washed off by rain or dropped to the ground as the leaves or branches fall (Nowak et al., 2006). Therefore trees can be considered as a temporary storage of pollution. The rate at which pollutants are removed is dependent on the species and size of the tree. Nowak et al. (2010) estimated that Chicago’s trees are responsible for the removal of 888 tons of air pollution annually, which has equates to a value of $6.4 million (based on national median externality costs associated with pollutants).

The deficiencies of traditional dendrometry have encouraged the development of new terrestrial remote sensing technologies for forest resources. Structure-from-Motion with Multi-view Stereo-photogrammetry (SfM-MVS) is one such technology, and its applicability for estimating tree metrics will form the basis of this thesis. Though it is still a relatively recent development, SfM-MVS may yet prove to be a superior method to traditional dendrometry.

1.4 Thesis Structure

This thesis is divided into four chapters; a comprehensive literature review immediately follows this introduction chapter, while the final two chapters present the results of the research itself. Trees used in the study were divided into two subsamples (nursery trees and landscape trees) based on the different research aims for each and therefore the results corresponding to nursery and landscape trees have been divided into separate chapters.

(17)

7

Chapter Two

Remote Sensing in Forestry

Remote sensing refers to the ability to obtain information of something from a remote location without making physical contact. It includes technologies such as radar, laser scanning, satellite imaging and digital photogrammetry, all of which are highly useful for analysing tree metrics and forest resources (Kankare et al., 2013a; Fernández-Sarría et al., 2013). Remote sensing is so useful due to the versatility of the collected data and the numerous ways the data is able to be interpreted, i.e. “a picture paints a thousand words”. The practicalities of analysing the vast land areas associated with commercial forest plantations (thousands of ha) mean that aerial platforms often provide the highest efficiency in terms of the area able to be covered at once and thus aerial laser scanning and digital photogrammetry are the primary remote-sensing methods with which forest resources are analysed in modern times (West, 2009; Kaartinen et al., 2012).

These methods have been shown to provide results at least as accurate as traditional dendrometry (Næsset, 2004; Melkas et al., 2008), the latter of which have been in use for over a century (West, 2009). However, no single remote sensing method is suited to all forestry applications– each has its own drawbacks and limitations depending on the nature of the research. Large scale analysis often requires plot-size ground validation, meaning supplementary field mensuration is still relied upon. Not all applications that require acquisition of tree metrics involve such large spatial areas, so remote sensing methods applicable to small-scale analysis (i.e. at plot and individual tree scale) are also important.

Remote sensing of forest resources began with hand-held optical dendrometers in the early 19th century (Drew et al., 2009) and advances in the technology of the sensors and the platforms on which they are based on has seen progression to modern day instruments. Hand-held devices are still commonly used, due to the necessity of field mensuration for many applications, though there is an increasing reliance on automatic extraction of forest metrics from modern digital datasets. Continuing advances mean forest mensuration, particularly for inventory, is trending towards full automation and manual field measurement may eventually become redundant (West, 2009). The following literature review provides an introduction to relevant remote sensing technologies and will help to identify study methods and also thresholds for success, against which the SfM-MVS technology used in this thesis, can be compared. Particular attention is given to laser scanning, optical dendrometry and photogrammetry techniques. The current state of SfM-MVS use for tree mensuration is also summarised and gaps in the research identified.

2.1 Laser Scanning

Laser scanning was first applied in forestry several decades ago (Nelson et al., 1984) and since then advances in the technology and associated methodologies have led to widespread application for mensuration of forest resources. It is now the preferred tool for making 3D measurements of forest

(18)

8 structure and analysing forest resources (Dandois & Ellis, 2010; White et al., 2013). Laser scanning, namely LiDAR (a portmanteau of ‘light’ and ‘radar’ or an acronym for Light Detection and Ranging) has overcome many of the limitations and reduced the errors associated with traditional field mensuration. It has proven to provide accurate estimates of a range of metrics (e.g. height, DBH, stem volume) across a range of coniferous species (Næsset, 1997; Lim et al., 2003; Kaartinen et al., 2012) and is becoming cheaper and more widely available with time (Henning & Radtke, 2006; Tanhuanpää et al., 2013). LiDAR works by determining distance to an object by measuring properties of light reflected off it. The distance to the object or surface is calculated by measuring the time delay between the transmission of the laser and the reception of the reflected signal (Lim

et al., 2003; Li et al., 2012). LiDAR pulses are able to penetrate forest canopies offering the ability to capture detailed information on both horizontal and vertical forest structure (Moorthy et al., 2011). Aerial platforms are most commonly used for inventory as they offer the ability to study large areas efficiently, though terrestrial based scanners are frequently used for research and small scale analysis.

2.1.1 Remote Sensing in Forestry - Aerial Laser Scanning

Aerial laser scanning (ALS) is being increasingly used for commercial forest inventories (Kaartinen et al., 2012) and urban tree inventories (Saarinen et al., 2014). It has been the subject of extensive research over the last decade and is used for a range of applications in mensuration of forest resources such as tree height, basal area, plot characteristics, timber volume, biomass, canopy structure and species type (e.g. Næsset, 1997; 2002; 2004; Lefsky, et al., 1999; Li et al., 2012; Kankare et al., 2013b; Vastaranta et al., 2013). ALS is used almost exclusively for analysis on a large spatial scale as the costs and logistics involved render it an inefficient tool for mensuration of small study areas. In many cases ALS data is available through external sources; as of 2011 it was estimated that 30% of the United States land area had ALS coverage of some form and some provinces in Canada are gradually analysing more of their land area (White et al., 2013). In some Scandinavian countries, ALS is becoming the preferred tool for monitoring national forest inventories (Kaartinen et al., 2012) and several other European countries are in the process of acquiring nation-wide ALS coverage (White et al., 2013).

The weight of the equipment required in ALS means it is operated from small aircraft, flying at altitudes between 400 m (e.g. Vastaranta et al., 2012) and 700 m (e.g. Næsset, 2002); the swath width, or footprint of the laser scan depends on the flying altitude. Both small-footprint (e.g. Næsset, 1997; 2002) and large-footprint (e.g. Lefsky et al., 1999) have been used, and 50% overlap between swaths is targeted in order to ensure total coverage of the study area. ALS systems can generally be categorised as either discrete or full waveform. Full waveform laser scanners are the most commonly used ALS system (Wulder et al., 2008) and are capable of penetrating several layers of forest canopy, providing estimates of vertical vegetation distribution and forest structure as well as ground elevation (Dassot et al., 2011; White et al., 2013). ALS data is used to produce 3D point cloud models based on all features intercepted by laser pulses. From this, information on tree dimension and forest structure can be extracted using the various proprietary software programmes that are available for post-processing of ALS data. This process is mostly automated

(19)

9 and the refinement of algorithms has been on going over the last decade. The point clouds can be geo-referenced, allowing the creation of digital elevation, surface and canopy models.

Though ALS has proven itself to be a highly accurate method of forest mensuration, it isn’t without its limitations. As laser pulses are transmitted from above, ALS can directly measure tree height, but relies on allometric relationships and statistical inference for other metrics required for inventory calculations (e.g. DBH, basal area). It is the major uncertainty around the predicted metrics that leads to inaccuracies when extrapolated to plot-scale volume measurement (Vastaranta et al., 2012). ALS is a very expensive tool and while its cost has reduced over time, the high cost presents a major obstacle to application. Though it has proven to be cost-effective and cheaper than traditional field mensuration for commercial forest plantations where the area of analysis often spans thousands of hectares, (Holopainen & Talvitie, 2006; Kaartinen et al., 2012; Jakubowski et al., 2013) it is usually totally unaffordable for small-scale analysis or non-commercial purposes. When applied over large areas it generally costs $0.05-0.2 per ha, though typical commercial ALS often costs a minimum of $20,000 per flight, regardless of study area size (Dandois & Ellis, 2013). Javernick et al. (2014) was quoted $28,000 for an ALS survey covering 1.6 x 0.65 km2, with $26,000 for repeated surveys. Such high cost means ALS datasets are typically acquired at low temporal frequency (Eitel et al., 2013). Generally higher laser pulse density is sought for forestry applications due to the higher precision that can be achieved, though this results in an increase in cost (Baltsavias, 1999). The cost is ultimately determined by the duration that the aircraft is in flight (Jakubowski et al., 2013).

2.1.2 Remote Sensing in Forestry - Terrestrial Laser Scanning

Terrestrial laser scanning (TLS) or ground-based scanning has been used in many previous studies to measure such forest metrics as tree height and stem diameter (Watt & Donoghue, 2005; Henning & Radtke, 2006; Maas et al., 2008; Vastaranta et al., 2009; Eitel et al., 2013; Liang et al., 2014a), tree location (Liang et al., 2012), stem count density, timber volume (Hopkinson et al., 2004), leaf area index (Strahler et al., 2008), crown volume (Moorthy et al., 2011) and biomass (Kankare et al., 2013b; Yu et al., 2013; Skowronski et al., 2014). As they are ground-based, TLS systems are good at capturing vertical vegetation structure and distribution (Moorthy et al., 2011) and are capable of directly measuring tree metrics to millimetre accuracy (van Leeuwen & Nieuwenhuis, 2010; James & Robson, 2012) rather than relying on allometric relationships to derive metrics like ALS. Being positioned on the ground means TLS is more suitable for applications involving only a small sample area (a few tens of metres, Holopainen et al., 2013) and where small detail changes over time are of interest (Hopkinson et al., 2004; Kankare et al., 2013b). It also means they are closer to the target objects, resulting in higher point cloud densities. In a sense TLS fills the niche between ALS and traditional dendrometry (Maas et al., 2008; Dassot et al., 2011). TLS systems are typically mounted on some form of tripod with a pan and tilt rotator on top. The scanning range of a midrange TLS system is 800 m (Kankare et al., 2013a) though for most applications the required range is less than 100 m. Algorithms are typically used to autonomously identify trees in the LiDAR point cloud in the various post-processing software that are available. Mobile laser scanning (MLS) is an adaptation of TLS which involves a laser scanner operating from a moving vehicle or some other form of mobile platform (Lin et al., 2010; Holopainen et al., 2013). In

(20)

10 many ways it faces the same issues as TLS, with the only real advantage of MLS over TLS being that due to being mounted on a vehicle it is very mobile, though it is limited to places with vehicle access.

Estimation of tree heights has not been as successful as stem diameter, especially compared with ALS data (van Leeuwen & Nieuwenhuis, 2010), mainly due to occlusion from the canopy. In some studies canopy occlusion has meant an accuracy of only 3-5 m could be achieved (e.g. Maas et al., 2008). However, some studies have still reported relatively high accuracy of <2 m or <10% (e.g. Hopkinson et al., 2004; Eitel et al., 2013; Kankare et al., 2013b).

Retrieval of stem diameter and DBH values has been particularly successful (e.g. Kankare et al., 2013; Liang et al., 2014a) though generally only in the lower stems where occlusion from branches is not an issue, (Henning & Radtke, 2006; Maas et al., 2008). The DBH accuracies achieved have typically been 1-2 cm (e.g. Hopkinson et al., 2004) though some of the more recent studies have produced accuracies to within >1 cm (e.g. Liang et al., 2014a).

There have been fewer studies utilising TLS to measure tree volume than linear metrics, and those that have mainly focused on stem volume. Remaining branch volume and total tree volume have also been estimated. The results have been mixed, with some studies producing estimates containing <10% error (e.g. Hopkinson et al., 2004), while others have been much less accurate with error >15% (e.g. Kankare et al., 2013). Remaining branches have not been estimated by many studies but the few that have produced quite inaccurate results (e.g. Dassot et al., 2012; Kankare et al., 2013).

Examples of the accuracy achieved in previous TLS studies for height, DBH and volume can be seen in Tables 2.1, 2.2 and 2.3). Not all studies report full statistics so they are only included where provided.

Table 2.1: Examples of the accuracies of height estimates achieved in previous TLS studies.

Hopkinson et al. (2004) Maas et al. (2008) Moorthy et al. (2011) Eitel et al. (2013) Kankare et al. (2013b) RMSE (%) 7 - - - 8.1 RMSE (m) 1.5 4.6 0.21 0.01-0.1 1.5 R² - - - 0.98-1.0 -

Table 2.2: Examples of the accuracies of DBH estimates achieved in previous TLS studies.

Hopkinson et al. (2004) Strahler et al. (2008) Eitel et al. (2013) Kankare et al. (2013b) Liang et al. (2014) RMSE (%) - - - 7.1 4.2 RMSE (cm) 1 - 2.2 1.5 0.8 R² - 0.34-0.66 0.99 - -

(21)

11 Table 2.3: Accuracy of tree volume estimates achieved in previous TLS studies (all values are RMSE %). Hopkinson et al. (2004) Dassot et al. (2012) Kankare et al. (2013b) Yu et al. (2013) Liang et al. (2014a) Stem volume 7 <10 15.3 - -

Remaining branch volume - 10-30 23.4-31.1 - -

Total volume 7 - 16.7 - 9.5

Above ground biomass - - 11.9-12.9 12.5 -

While TLS provides advantages to ALS, there are still disadvantages that may render it unsuitable for certain applications or users. Occlusion of trees is a significant issue; in forests with high tree density the usefulness of TLS decreases due to occlusion, especially in the upper canopy (Watt & Donoghue, 2005; van Leeuwen & Nieuwenhuis, 2010). Laser pulses cannot penetrate beyond the trees that are immediately in front of them and leaves and undergrowth can limit the field of view, preventing distant trees from being identified (Maas et al., 2008). Occlusion resulted in Moskal and Zheng (2012) reporting a capture rate of only 18% of the total tree volume for the plot. Strahler et al. (2008) only managed to find 40% of the total trees that were manually identified in their study; of these, 33% were either partly or fully occluded. For plot-scale analysis multiple scans from different positions may be required to help overcome occlusion of individual trees (Moskal & Zheng, 2012; Kankare et al., 2013b). TLS may even be better suited for analysis of individual trees (Watt & Donoghue, 2005) as their short and limited working range reduces their practicality for large-scale inventory analysis (van Leeuwen & Nieuwenhuis, 2010).

Forest mensuration using TLS is subject to the same accessibility issues as traditional dendrometry, perhaps even more so as bulkier equipment is required. It is expensive, labour-intensive and time consuming to conduct repeated surveys (Moskal & Zheng, 2012) and the equipment required can be heavy, with some TLS systems being as heavy as 13 kg (e.g. Maas et al., 2008, Moorthy et al., 2008, Kankare et al., 2013b and Liang et al., 2014a) and some even weighing upwards of 20 kg (Dassot et al., 2011). There are lightweight (<5 kg) models available (e.g. Eitel et al., 2013) though a tripod, pan-tilt unit, battery or generator and field computer are also necessary items in the field. TLS systems are costly; entry-level systems can cost tens of thousands of dollars (Eitel et al., 2013; Liang et al., 2014a) and high-end systems can be upwards of $500,000. Less advanced rangefinder TLS systems such as that used by Parker et al. (2004) can cost up to $8000 and even the laser system assembled by Eitel et al. (2013) using ‘off-the-shelf’ components cost nearly $12,000. Though 2D LiDAR sensors can offer a cheaper alternative to 3D sensors the cost is also relatively high (Rosell et al., 2009). Some expertise is necessary to operate them and the time required for data acquisition as well as post-processing should also be considered. The high resolution scans can run for up to two hours and the amount of data obtained is enormous (West, 2009).

2.2 Optical Dendrometry and Photogrammetry

Though laser scanning is the leading tool for mensuration of forestry resources and tree metrics in modern times, remote sensing in the field had far more modest beginnings. Optical dendrometry

(22)

12 refers to mensuration of tree height and stem diameter made with sight which does not require any contact with the tree (Clark et al., 2000b). Optical dendrometry includes hand-held instruments like optical callipers, rangefinder dendrometers, relascopes, clinometers and cameras, all of which have been shown to provide accurate measurements of diameter (Parker, 1997). These instruments were developed to fill the niche afforded by their ability to remotely capture information from parts of the tree that could not be physically accessed, e.g. the upper stems of tall trees. Development of new instruments is on-going (Liang et al., 2014b) and laser relascopes, hypsometers and digital callipers are among the most recent. These instruments still rely on manual retrieval and some are limited in what parts of the tree they can measure (i.e. stem diameter or height). Traditionally, hand-held or ground-based cameras have not been commonly used for tree mensuration but digital cameras may hold the most potential as field instruments because as well as being able to capture information quickly, they are also able to store it indefinitely, meaning data can be extracted and used in the future.

Photogrammetry is the science of obtaining measurements of objects and surfaces from photographic images. When the scale of an image is known, distance measurements can be made between two points on a plane parallel to that of the image; this is the simplest example of a two-dimensional calculation, though 2D photogrammetric measurements are relatively straight forward compared with 3D measurements made with multiple images. A basic example in two overlapping images, the 3D location relative to the camera can be calculated for a set of given pixels that appear in both images (Bemis et al., 2014). Photogrammetry is useful for situations where physical measurements are difficult or impossible to obtain, due to their inaccessibility or the fact that the object of interest doesn’t exist anymore (but photos of it remain). It is also useful for material that can be ‘easily transformed’, e.g. water and loose sediment (Agisoft LLC, 2012).

One of the first to use a hand-held camera as a form of dendrometry was Marsh (1952), who took photos of tree diameters and reported errors of 63.5 mm and 20.3 mm for horizontal and oblique photos respectively. The results from this study were improved by Ashley and Roger (1969) who designed a frame that reduced problems associated with scale and orientation of the setup. This study produced errors of 7.6 mm for diameter measurements in their laboratory setting. Bradshaw (1972) was another to use a camera to measure the diameter at 26 points along a single stem, which ranged in width from 30 to 76 cm. An accuracy of 9.9 mm was reported. Crosby et al. (1983) used a handheld camera to photograph pine trees in order to extract diameter and height information. The study produced relatively accurate estimates of tree diameters, with error of <2% for trees with diameters less than 50 cm. Clark et al. (2000b) trialled the ability of a commercially available digital camera to collect the heights and diameters of trees and from these measurements, estimate their subsequent stem volumes. In this study the camera proved itself to be a promising instrument in assessing forest metrics, with stem volumes derived from photographs found to be within 8% of those calculated using physical measurement techniques. Examples of the accuracy achieved in previous photogrammetry for height, DBH and volume can be seen in Tables 2.1, 2.2 and 2.3). Not all studies report full statistics so they are only included where provided.

(23)

13 Table 2.4: Examples of the accuracies achieved estimates of stem diameter, height and volume in previous photogrammetric studies.

Ashley & Roger (1969) Bradshaw (1972) Crosby et al. (1983) Clark et al. (2000b) Dean (2003) Stem Diameter 0.76 cm 0.99 cm 2% 4 cm - Height - - - - - Stem Volume - - - 8% 10%

The use of cameras has presented challenges for those attempting to use them as a tool for dendrometry. In each of the aforementioned studies analogue cameras were used. The cost of film and the time required for its development, and issues with exposure meant it was a tool really only suited to situations where field measurements of a tree would have otherwise been difficult to obtain. The geometry of the camera setup also caused error in the measurements, with the oblique line of sight and 2D nature of the photos contributing to an underestimation of the actual stem diameter. Forests can provide difficult conditions in which to use cameras as a means of mensuration, as undergrowth and other trees can prevent a clear line of sight (Clark et al., 2000b). The inability to capture an entire tree in one image or even view the very tops of trees within a forest setting also meant that cameras could not be used for tree height measurements. For single tree diameter measurements (i.e. two-dimensional) other forms of optical dendrometry are just as capable as the camera.

2.2.1 Remote Sensing in Forestry - Stereo-photogrammetry

Since the invention of the digital camera in 1975 advances in digital image technology and reduction in cost has facilitated the emergence of digital photogrammetry, based on stereo image pairs (stereo-photogrammetry), as an effective tool for 3D topographical modelling (Clark et al., 2000b; Westoby et al., 2012; Vastaranta et al., 2013). The development of aerial digital cameras enabled easy acquisition of collections of overlapping photos and advances in computing power have allowed development of complex algorithms capable of matching large collections of images (Vastaranta et al., 2013; White et al., 2013). The refinement of these algorithms designed for automated extraction information from digital images has been the subject of intense research and they are being continuously developed (Baltsavias, 1999). Forestry was one of the ﬁrst fields to use aerial photographs to produce 3D models of forests and canopies (Rosell et al., 2009).

For 3D reconstruction of a target object, multiple overlapping 2D images from two different perspectives are required. This is termed stereoscopic viewing, and when understanding the concept it is helpful to consider the way human eyesight works. Each eye is located in a different position and therefore has a slightly different view of a given target object, meaning the brain receives a different image from each respective eye and must interpret. Triangulation between both eyes and objects in the field of view allows for the generation of depth perception. For stereo-photogrammetry, reference points are identified across a range of these images. A line of sight extends from the camera viewpoint to a reference point in each image and triangulation is used to find the intersection of the two rays so the 3D dimensional position can be determined (White et al., 2013).

(24)

14 Stereo-photogrammetry is based on the motion parallax principle, where the apparent motion of stationary objects relative to their background between viewing positions allows for estimation of their relative distance (White et al., 2013), i.e. nearby objects will appear to move while far away objects will appear to remain still.

High resolution digital stereo imagery (DSI) has been used to generate digital canopy models with accuracy close to that of laser scanning (Baltsavias, 2008; Vastaranta et al., 2013), measure tree and forest height characteristics and has proved to be suitable for tracking tree height growth and long-term canopy dynamics (Lisein et al., 2013; Vastaranta et al., 2013). Virtually all DSI for forest mensuration has been from aerial platforms and it is probably best suited to providing height measurements as photography cannot penetrate closed-canopy forests to provide information on the ground surface or forest structure beneath (Lisein et al., 2013). When coupled with an ALS-derived digital terrain model, a canopy height model can be produced through subtraction of ground elevations from canopy elevations. The quality of these height measurements are determined by the quality of the DEM that is used (Vastaranta et al., 2013).

Once LiDAR data has been used to produce a DEM, with the assumption that the ground elevation remains consistent in a given area, monitoring of future forest structural changes can be done using only recurrent aerial photogrammetry (St-Onge et al., 2004). The idea has been suggested that ALS data be acquired only periodically, e.g. every 10 years, and in between DSI used to provide more regular updates (White et al., 2013). Satellites may provide a potential means of imagery with which to produce surface models, though at present satellite imagery is too coarse to produce DEMs of comparable resolution to low altitude platforms (Westoby et al., 2012; Kankare et al., 2013a).

The general practical requirements of DSI place restrictions on its use; these include the need for images to be as parallel as possible and contain 60% overlap with each other, as well as the need for ground-control points of known co-ordinates for geo-ferencing (James & Robson, 2012). There are a range of issues that affect the image quality of DSI. Image-matching of vegetation is challenging in the fact that the visual texture is often very repetitive and smaller trees are often occluded (Lisein et al., 2013). The cost of DSI is roughly one-half to one-third that of ALS data, though it varies constantly and is highly dependent on the nature of the work being undertaken (White et al., 2013).

2.3 Remote Sensing in Forestry - Structure from Motion and Multi-View Stereo-photogrammetry Advances in computer vision technology have led to the availability of a computer-based photogrammetric method called Structure-from-Motion (SfM), which when paired with Multi-View Stereo-photogrammetry (MVS) provides a semi-automated, low-cost option for generating high resolution 3D point cloud models from collections of 2D photographic images (Westoby et al., 2012; Javernick et al., 2014).

SfM reconstruction relies on changes in the position of features known as ‘keypoints’, from one image to another. As each image is captured from a different spatial position, the perspective of the scene is also different and therefore is represented by a different arrangement of pixels, and

(25)

15 therefore keypoints, in each image. Substantial overlap between successive images is necessary in order to recognise keypoints and calculate camera positions and scene geometry.

Though the mechanics are similar to traditional stereo-photogrammetry, SfM differs fundamentally in the fact that it uses robust algorithms to solve 3D camera location and orientation and scene geometry and removes the need for a network of targets to be established prior to image capture (Morgenroth & Gomez, 2014). The point cloud models have RGB colour attributes (Dandois & Ellis, 2013) and are of comparable quality to those made with laser scanning (Westoby et al., 2012). The term ‘motion’ in SfM stems from these images which are captured from different locations and when viewed in sequence create a sense of movement (Snavely et al., 2008).

SfM has origins in the field of computer vision in 1979 (Ullman, 1979) and its development has continued in the following decades, with advances in image-matching algorithms (Snavely et al., 2010). SfM-MVS is a relatively unused technology in the fields of geoscience and natural resources (Westoby et al., 2012) but it has been the focus of an increasing number of studies in recent year and has been used successfully in the fields of geomorphology (e.g. James & Robson, 2012; Westoby

et al., 2012; Bemis et al., 2014; Gienko & Terry, 2014 and Javernick et al., 2014), agriculture (e.g. Bendig et al., 2013) and to a lesser extent, forestry (e.g. Dandois & Ellis, 2010; 2013; Liang et al., 2014a; 2014b; Morgenroth & Gomez, 2014) and has been able to accurately model objects and surfaces in spatial scales ranging from centimetres to kilometres (Dandois & Ellis, 2010; James & Robson, 2012; Javernick et al., 2014; Gomez et al., 2015).

The benefits of using the camera-based method of SfM-MVS as an alternative to laser scanning methods for 3D modelling are centred on the low cost and low bulk associated with the equipment required (James & Robson, 2012; Dandois & Ellis, 2013).

2.3.1 The SfM-MVS Workflow

There are several software packages for processing SfM data, some of which are freely available. Some are browser-based (e.g. Microsoft's Photosynth) while others are desktop based (e.g. Bundler Photogrammetry Package, SFM toolkit, VisualSFM and PhotoScan). Little expertise is required to use these. Image processing and camera calibration are automated and few control points are required (James & Robson, 2012).

The reconstruction of SfM-MVS point clouds involves three steps. The first step is the SfM process, in which a series of algorithms identify keypoints in a collection of overlapping digital images. Each keypoint is represented by a distinctive arrangement of pixels that are likely to be recognised in other images (James & Robson, 2012). The SfM-MVS method requires substantial overlap between photos to be successful (50%) so the more images and the more overlap between each image in a series, the more keypoints will be matched between images and the more successful 3D reconstruction will be. Higher resolution images will require more processing time due to the larger number of pixels involved though if necessary these can be downsized before they are uploaded to the software.

In PhotoScan Professional by Agisoft, this process is fully automated (Dandois & Ellis, 2013). PhotoScan uses its own proprietary algorithms, which while similar to other commonly used

Estimation of individual tree metrics using structure-from-motion photogrammetry.